From bae382dc5cd20a6be342131692829263104cb83b Mon Sep 17 00:00:00 2001 From: RomirJ Date: Wed, 10 Jun 2026 22:19:27 -0700 Subject: [PATCH] =?UTF-8?q?docs:=20correct=20README=20parity=20claim=20?= =?UTF-8?q?=E2=80=94=204=20verified=20families,=20not=206?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Audit §3.3 (exporter-parity investigation, Claim 4). README:459 claimed "cos=+1.0000000 on six VLA families" including DreamZero and OpenVLA. But: - DreamZero export is config-only — exporters/dreamzero.py:91 returns status="config_only"; the DiT ONNX export is explicitly deferred (:94,108). - OpenVLA has no Tether ONNX path — exporters/openvla.py:92,99 raise NotImplementedError (points users at optimum-cli + a postprocess helper). Only SmolVLA, pi0, pi0.5, and GR00T N1.6 have a verified cos=1.0 Tether ONNX export (matching the hero paragraph at line 14, which already says "ALL four major open VLAs", and the parity table). Corrected line 459 to claim four verified families and describe DreamZero/OpenVLA accurately as supported-but- not-yet-ONNX-verified. Co-Authored-By: Claude Opus 4.7 (1M context) --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index 6ca4109..c8492cb 100644 --- a/README.md +++ b/README.md @@ -456,7 +456,7 @@ Every response surfaces telemetry from each enabled wedge (`guard_clamped`, `gua ## What Tether is and isn't -**Is:** the deployment layer between a trained VLA and a real robot. Cross-framework export verified at cos=+1.0000000 on six VLA families — SmolVLA + pi0 + pi0.5 (flow-matching, num_steps=10) + GR00T N1.6 (DDPM DiT, num_steps=4, **with Eagle 2.5 VL backbone producing live image+language KV**) + DreamZero (world-action model, joint video + action diffusion) + OpenVLA (shim) — plus a composable runtime (serve + safety + turbo + split), edge-first design targeting Jetson + desktop NVIDIA GPUs. +**Is:** the deployment layer between a trained VLA and a real robot. Cross-framework export verified at cos=+1.0000000 on four VLA families — SmolVLA + pi0 + pi0.5 (flow-matching, num_steps=10) + GR00T N1.6 (DDPM DiT, num_steps=4, **with Eagle 2.5 VL backbone producing live image+language KV**). DreamZero (world-action model — config + PyTorch runtime today, DiT ONNX in progress) and OpenVLA (optimum-cli shim + postprocess helper) are supported but not yet numerically verified through a Tether ONNX export. Plus a composable runtime (serve + safety + turbo + split), edge-first design targeting Jetson + desktop NVIDIA GPUs. **Isn't:** a training framework (PyTorch/JAX own that) or a cloud inference provider (vLLM/Baseten own that). Tether's moat is the deployment toolchain: cross-framework ONNX with verified numerical parity, composable safety wedges, ROS2 + Docker + HTTP serving, and a deterministic export receipt (`VERIFICATION.md`) your QA team can audit.